Condition on Discriminator

[tatajack]

Hi,

From page 5: To customize GANs for this task, one could condition on the given context information; i.e., the mask Mˆ ⊙ x. However, conditional GANs don’t train easily for context prediction task as the adversarial dis- criminator D easily exploits the perceptual discontinuity in generated regions and the original context to easily classify predicted versus real samples.

Does this means that we only rely on the "L1 reconstruction loss" to ensure the model to take the (outside) condition part into consideration? Since the discriminator only sees the inpainted region, it has no additional information to work on.

Looking forward to your reply! Thanks